Multi-line bogie fault diagnosis method, device and equipment based on federated learning and storage medium

Abstract

The invention provides a federated learning-based multi-line bogie fault diagnosis method, apparatus and device, and a storage medium. Relates to the technical field of fault diagnosis and federated learning. According to the method, through a step-by-step federated distillation architecture and by means of model knowledge of other lines, the diagnosis performance of a small sample line model is improved on the premise of ensuring data security; for a plurality of fault diagnosis tasks in the architecture, the loss is decoupled into TCKD and NCKD and further reconstructed, and the loss between a target class and a non-target class is balanced; for a plurality of distillation losses of multi-task output, a multi-loss weight adaptive adjustment strategy is provided, and efficient training is realized by cooperating with a plurality of distillation tasks; each line shares a basic shallow network based on model migration, and a personalized deep network is customized at the same time, so that the adaptability of a fault diagnosis model is improved, and the learning cost of the proposed architecture is reduced. According to the method, the fault modeling precision of a small sample line can be improved, the privacy security of multi-party data is ensured, and a feasible solution is provided for implementing federated learning among users with unbalanced data in an industrial scene.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis and federated learning, and in particular to a multi-line bogie fault diagnosis method, apparatus, device and storage medium based on federated learning. Background Art

[0002] As an important tool for railway transportation, high-speed trains must first ensure their safe operation. There are many factors that affect the safe operation of trains, one of the most important factors being the operation and maintenance of bogies. As one of the most important components in the structure of high-speed trains, bogies are mainly composed of wheelsets, bogie frames, primary suspensions, secondary suspensions, and basic braking equipment. It has the characteristics of complex mechanical structure, large number of components, and diverse operating conditions. While providing train power, it also bears the weight of the vehicle body, providing important guarantees for the safe and stable operation of the train. Since high-speed trains are affected by different operating environments and conditions such as long and long slopes, proportion of bridges and tunnels, and climatic conditions, as the high-speed trains run on track for an increasing amount of time, these influences often lead to a series of problems such as accelerated wheel wear, increased performance degradation of the primary and secondary suspension systems, and intensified impact vibration of the vehicle body and bogie system through the direct excitation of the bogie components and the indirect excitation coupling of the track. As a highly complex and highly coupled system, the failure of any component of the bogie system may cause a chain reaction, affecting the reliability of the entire vehicle. Therefore, accurate bogie fault diagnosis is crucial to ensure the safe and stable operation of high-speed railways.

[0003] The performance of fault diagnosis algorithms based on deep learning depends on a large amount of high-quality labeled data. However, in actual train operation and maintenance sites, it is often difficult to manually annotate data in detail, making it difficult to obtain high-quality labeled data, resulting in insufficient individual data, unbalanced sample types, incomplete information dimensions, and difficulty in establishing a complete model alone.

[0004] At the same time, in actual industrial scenarios, due to different operating environments of railways and different road conditions, the bogies of high-speed trains on different operating lines usually contain characteristic data that reflect their common faults, and the fault data is small-scale and fragmented. The model obtained by training only on a single line is only applicable to this line and has poor generalization ability. Therefore, if the bogie data of different lines are collected for unified training, a model with very strong generalization ability will be obtained. In theory, the model can accurately identify all fault conditions of the training line.

[0005] However, if the data of high-speed train bogies on multiple different lines are directly aggregated, the economic cost will be greatly increased. In addition, for some fields with high data sensitivity and easy to leak, such as finance and medical care, data is usually strictly protected and cannot leave the local storage device. At the same time, there may be data monopoly between enterprises, which makes data sharing and cooperation difficult to implement. In particular, data security issues are now receiving more and more attention, and national policies to protect data privacy have also been introduced. This makes data-driven methods have certain limitations in diagnosing faults in actual industrial environments. Therefore, how to effectively mine the intelligence contained in the data without infringing the privacy and security of the data itself is a difficult problem that needs to be solved nowadays.

[0006] A feasible idea is to jointly model multiple lines with different operating environments and working conditions based on federated learning. By migrating and reusing model knowledge between multiple lines and ensuring data privacy and security, mutual benefit and win-win results can be achieved. However, in the actual operation and maintenance of high-speed trains, due to the limitations of the actual equipment operating status and data collection implementation, when multiple lines are deployed according to the distributed learning architecture of federated learning, the high-speed train bogie data of the line often presents the characteristics of non-independent and identically distributed. The imbalance of data will lead to the imbalance of the line models of all parties, which is not conducive to the cooperation of multiple parties to establish a model with strong generalization ability, and will eventually seriously affect the training and detection of the diagnostic model.

[0007] In particular, in the current classic federated learning framework, each line has an equal status with the same weight, and multiple local models are averaged during model aggregation. However, lines with a small amount of data may also contain more valuable fault information, and direct aggregation of the model may mask the differences in their own data. The model drift of small sample lines causes each client to be unable to effectively learn information from other clients. Therefore, when multi-party data is unbalanced, giving multi-party lines the same priority in the architectural process may not have a positive effect on model optimization. Innovating the existing model aggregation process will improve the overall training efficiency of federated learning in some scenarios.

[0008] Secondly, since federated learning combines multiple data holders to build models, it often faces the problem of high computing overhead, resulting in low training efficiency and difficulty in implementation in actual scenarios.

[0009] From the above background, it can be clearly seen that when implementing federated modeling for multi-line high-speed trains, how to efficiently and accurately realize fault diagnosis under the federated learning framework when the multi-line bogie data is unbalanced has become an urgent problem to be solved. Summary of the invention

[0010] The present application provides a multi-line bogie fault diagnosis method, apparatus, device and storage medium based on federated learning to improve the existing federated learning framework and propose a new framework for data imbalance scenarios, thereby improving the model accuracy and generalization ability.

[0011] In a first aspect, the present application provides a multi-line bogie fault diagnosis method based on federated learning, comprising:

[0012] Sort multiple lines according to the amount of data, and obtain a multi-line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, where Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, n≥3;

[0013] Line Client1 constructs a deep neural network locally, and trains the deep neural network using Client1 local data, freezes the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmits the shallow frozen network and Client1 model parameters to the central server;

[0014] Line Client2 trains the Client2 model using the Client2 local data, and transmits the trained Client2 model parameters to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module;

[0015] Line Clienti trains the Clienti model using Clienti local data, and transmits the trained Clienti model parameters to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

[0016] In a possible design, line Client2 uses Client2 local data to train the Client2 model, including:

[0017] The shallow frozen network received from the central server does not participate in training;

[0018] A distillation loss function is constructed for the target class and the non-target class using the model parameters of Client1, and knowledge distillation is performed based on the distillation loss function to complete the local model training of Client2.

[0019] In one possible design, the distillation loss function is determined by the following method:

[0020] The classification probability p based on the target class is calculated by the following formula t :

[0021]

[0022] Among them, C is the number of categories, t represents the target category, z j represents the logarithmic value of the jth class, z t represents the logarithm of the target class;

[0023] The classification probability p based on the non-target class is calculated by the following formula \t :

[0024]

[0025] Among them, z k Represents the logarithmic value of the non-target class;

[0026] After removing the target class, the classification probability of other classes for:

[0027]

[0028] The KL divergence is used to measure the distillation loss between lines T and S:

[0029]

[0030] Among them, KD represents distillation loss, KL represents KL divergence calculation, and p T represents the classification probability distribution of route T, p S represents the classification probability distribution of route S, represents the classification probability of target class t in line T, represents the classification probability of target class t in line S, represents the classification probability of non-target class i in line T, represents the classification probability of non-target class i in route S;

[0031] make The distillation loss function is expressed as:

[0032]

[0033] Among them, b represents the binary classification scenario, TCKD is the target class knowledge distillation, and NCKD is the non-target class knowledge distillation. is the classification probability of target class t in line T, is the sum of the classification probabilities of non-target classes in line T, is the classification probability of the remaining categories i after removing the target class t in line T, is the classification probability of the remaining categories i in line S after removing the target class t, is the sum of the classification probabilities of non-target classes in line T, is the sum of the classification probabilities of non-target classes in line S, is the classification probability distribution of line T excluding target class t, is the classification probability distribution of line S without target class t, b T and b S They represent the binary classification scenarios of route T and route S respectively;

[0034] An additional weight value is given to TCKD through the following formula:

[0035] DKD=α·TCKD+β·NCKD

[0036] Among them, DKD represents the coupled distillation loss (Decoupled Knowledge Distillation), α is the adjustment parameter for balancing TCKD and NCKD, and β is used to represent

[0037] In a possible design, line Clienti uses Clienti local data to train Clienti model, including:

[0038] The shallow frozen network received from the central server does not participate in training;

[0039] Obtain line model parameters with a higher priority than line Clienti from the central server;

[0040] A one-to-one corresponding distillation loss function is constructed between the Clienti model and each line model parameter having a higher priority than the line Clienti, and the weight of the distillation loss function is dynamically adjusted to obtain an overall loss function of the Clienti model, and the Clienti model is trained using the overall loss function.

[0041] In one possible design, the overall loss function is expressed as:

[0042]

[0043] Among them, for the i-th line in the federated task (i=1,2,L,N), represents the local multi-classification cross entropy loss function value, represents the bi-norm value of the DKD loss of the t-th distillation task, represents the average bi-norm value of DKD loss for all tasks, r trepresents the loss training completion of the t-th distillation task, λ is a hyperparameter that adjusts the local model training and the knowledge transfer from other parties, the item where 1-λ is located is used to realize the knowledge distillation between multiple lines, μ is a hyperparameter that controls the joint training rate between multiple tasks, and the item where it is located is used to balance the overall loss, and M is the total number of distillation tasks;

[0044] In each distillation task, Set to a fixed constant and do not participate in the back propagation of gradients during training;

[0045] The local cross entropy loss value of the i-th line is:

[0046]

[0047] Among them, C represents the number of categories, y k Indicates the true label corresponding category, Represents the probability predicted by the softmax function;

[0048] The loss training completion of the t-th distillation task r t for:

[0049] r t =L t / L avg

[0050] Among them, L t is the average cross entropy loss between lines, L avg is the average loss of all lines, r t The smaller the value, the higher the training completion.

[0051] In one possible design, before sorting the multiple lines according to the data amount, the method further includes:

[0052] Obtain normal and fault data of high-speed train bogies on different lines; wherein, the normal and fault data of high-speed train bogies include vibration signals under normal working conditions and vibration signals under fault conditions, and the fault conditions include AS air leakage, LD oil leakage, AD oil leakage, LD oil leakage + AD oil leakage, AS air leakage + LD oil leakage and AS air leakage + AD oil leakage.

[0053] In a second aspect, the present application provides a multi-line bogie fault diagnosis device based on federated learning, including a central server and multiple lines, wherein the central server is signal-connected to the multiple lines:

[0054] The central server is used to sort the multiple lines according to the amount of data, and obtain a multi-line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, wherein Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, n≥3;

[0055] Line Client1 is used to construct a deep neural network locally, and train the deep neural network using Client1 local data, freeze the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmit the shallow frozen network and Client1 model parameters to the central server;

[0056] Line Client2 is used to train the Client2 model using the local data of Client2, and the parameters of the Client2 model obtained by training are transmitted to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module;

[0057] Line Clienti is used to train the Clienti model using Clienti local data, and the trained Clienti model parameters are transmitted to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

[0058] In the third aspect, the present application provides an electronic device comprising: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the multi-line bogie fault diagnosis method based on federated learning as described in the first aspect and various possible designs of the first aspect.

[0059] In a fourth aspect, the present application provides a computer-readable storage medium, in which computer execution instructions are stored. When a processor executes the computer execution instructions, the multi-line bogie fault diagnosis method based on federated learning as described in the first aspect and various possible designs of the first aspect is implemented.

[0060] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the multi-line bogie fault diagnosis method based on federated learning as described in the first aspect and various possible designs of the first aspect.

[0061] The multi-line bogie fault diagnosis method, device, equipment and storage medium based on federated learning provided by the present application have the following beneficial technical effects:

[0062] 1. Aiming at the data imbalance problem in multi-line federated modeling, this application proposes a step-by-step federated distillation architecture based on knowledge distillation, which uses models of other lines to improve the fault diagnosis performance of small sample line models while ensuring data security.

[0063] 2. In each distillation task, this application decomposes the distillation loss into two parts, TCKD and NCKD, and then recouples them to flexibly exert the knowledge distillation's ability to characterize the target and non-target classes of bogie faults.

[0064] 3. This application proposes a loss weight adaptive adjustment strategy for multiple knowledge distillation tasks based on grad-norm to achieve efficient synchronous collaborative training among multiple tasks of inter-line bogie fault diagnosis.

[0065] 4. In the multi-party training of this architecture, the shallow model of the party with the highest priority is frozen and shared with other parties, reducing the cost of model training. At the same time, the fault diagnosis model of each line is configured with a personalized module, which improves the adaptability of the local model and balances the generalization and personalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0067] Figure 1 It is a schematic diagram of an application scenario in the prior art;

[0068] Figure 2 A flowchart of a multi-line bogie fault diagnosis method based on federated learning provided in an embodiment of the present application;

[0069] Figure 3 A diagram of common fault types of a bogie provided in an embodiment of the present application, wherein: (a) indicates air spring (AS) leakage, (b) indicates lateral damper (LD) oil leakage, and (c) indicates anti-yaw damper (AD) oil leakage;

[0070] Figure 4 A bogie fault diagnosis flow chart for four lines provided in an embodiment of the present application;

[0071] Figure 5 A schematic diagram of constructing a local personalized model for high-speed train bogie fault diagnosis on each line provided in an embodiment of the present application;

[0072] Figure 6 The overall framework diagram of step-by-step federal distillation using four lines as an example provided in the embodiment of the present application;

[0073] Figure 7 A schematic diagram of the structure of a multi-line bogie fault diagnosis device based on federated learning provided in an embodiment of the present application;

[0074] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0075] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0076] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0077] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of information such as financial data or user data involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.

[0078] It should be noted that in the embodiments of the present application, certain software, components, models and other existing solutions in the industry may be mentioned, and they should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0079] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of information such as financial data or user data involved shall comply with the provisions of relevant laws and regulations and shall not violate public order and good morals.

[0080] It should be noted that in the embodiments of the present application, certain software, components, models and other existing solutions in the industry may be mentioned, and they should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0081] Figure 1Figure 2 is an application scenario diagram of a multi-line bogie fault diagnosis method based on federated learning. Figure 1 As shown, the application scenario includes a central server 101 and multiple lines 102, each line is represented by Client1, Client2, Clienti, Clientn; wherein n represents the total number of lines, the total number of lines is determined according to actual conditions, and it is at least two. The central server 101 is used to sort multiple lines according to the amount of data, and obtain a multi-line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, wherein Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, n≥3; line Client1 constructs a deep neural network locally, and uses Client1 local data to train the deep neural network, freezes the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmits the shallow frozen network and Client1 model parameters to the central server; line Client2 uses Client2 local data to train the Client2 model, and the trained Client2 model parameters are transmitted to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module; line Clienti uses Clienti local data to train the Clienti model, and the trained Clienti model parameters are transmitted to the central server; wherein the Clienti model includes the shallow frozen network received from the central server and the i-1th personalized module, i is a natural number and 3≤i≤n.

[0082] It should be noted that the line refers to the existing rail operation line, and each rail operation line includes a data processing unit. The process of knowledge distillation of each line as described above is executed by the data processing unit included in the corresponding rail operation line; the local data of each line is the normal and fault data of the train bogie collected by the line. Different levels of federated distillation structures are built in turn for each line according to the amount of data. For example, the line Client1 with the largest amount of data is responsible for transmitting the model information trained by itself to all other lines through the central server, and transferring the model knowledge through the knowledge distillation technology, which constitutes a first-level federated distillation structure. The line Client2 with the second largest amount of data conducts local model training, forming a second-level federated distillation structure. The line Client3 with the third largest amount of data conducts local model training, forming a third-level federated distillation structure. The line Client4 with the fourth largest amount of data conducts local model training, forming a fourth-level federated distillation structure. By analogy, an n-level federated distillation structure is realized, and the fault diagnosis performance of the small sample line model is improved under the premise of ensuring data security with the help of the models of other lines.

[0083] Figure 2 A flowchart of a multi-line bogie fault diagnosis method based on federated learning provided in an embodiment of the present application. Figure 2 As shown, the embodiment of the present application provides a multi-line bogie fault diagnosis method based on federated learning, the method comprising the following steps:

[0084] S210. Sort multiple lines according to the amount of data to obtain a multi-line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, where Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, where n≥3.

[0085] In this embodiment, n is 4, that is, taking 4 lines as an example, the multi-line sequence is represented as Client1, Client2, Client3 and Client4 in sequence.

[0086] Before step S201, this embodiment collects normal and fault data of high-speed train bogies on different lines through the SIMPACK platform, and then sorts multiple lines according to the data volume of normal and fault data of high-speed train bogies on different lines, so as to determine the priority of each line, so as to facilitate the design of the subsequent multi-level federal distillation structure.

[0087] Due to the limitations of actual industrial conditions, the cost of obtaining data through real high-speed train bogie fault experiments is very high. Therefore, the SIMPACK platform can effectively simulate and collect bogie fault data. The State Key Laboratory of Traction Power of Southwest Jiaotong University provided a nonlinear dynamic model of the CRH380A high-speed train system for the experiment. The entire train model has 62 degrees of freedom. In the vertical, lateral and longitudinal directions, sensors installed at different positions of the vehicle bogie obtained vibration signals, including acceleration and displacement of different components. The experiment recorded a total of 58-dimensional vibration signals.

[0088] The vibration data of the embodiment of the present application is generated by the track spectrum excitation of the Wuhan-Guangzhou Line, Zhengzhou-Xi'an Line, Beijing-Tianjin Line and Jiaoji Line in SIMPACK. The train bogie consists of two air springs (ASs), two lateral dampers (LDs) and four anti-snaking dampers (ADs), which are common faults of the bogie. The fault state is as follows Figure 3 As shown in the figure, they are AS air leakage, LD oil leakage and AD oil leakage. These phenomena indicate that the bogie components have been degraded to a certain extent, which will aggravate the vibration of the bogie and even affect the comfort and safety of the train. The experiment studied 7 common working conditions on each line, corresponding to labels 0, 1, 2, ..., 6, as shown in Table 1.

[0089] Table 1 Common working conditions of bogies

[0090] Label Working conditions 0 normal 1 LD 2 AD 3 AS 4 LD+AD 5 AS+LD 6 AS+AD

[0091] The vibration data sampling frequency is set to 243 Hz, and the train running speed is set to 200 km / h. For each working condition, a small amount of time-series vibration signals are collected, and the corresponding data volume corresponds to the priority of the line in the framework decreasing step by step. The specific details are shown in Table 2.

[0092] Table 2 Design of bogie data for different lines

[0093]

[0094] S220, line Client1 builds a deep neural network locally, and uses Client1 local data to train the deep neural network, freezes the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmits the shallow frozen network and Client1 model parameters to the central server.

[0095] Figure 4 A bogie fault diagnosis flowchart for four lines is provided in an embodiment of the present application.

[0096] like Figure 4 As shown, step S220 specifically includes:

[0097] S221 and Client1 build a deep neural network locally and only use their own local data to train the model.

[0098] S222: After freezing the shallow network of the model trained by Client1, encrypt the model parameters and upload them to the central server.

[0099] S223. The central server sends the model parameters of Client 1 to the remaining low priority lines (Client 2, Client 3, Client 4).

[0100] S230, line Client2 uses Client2 local data to train the Client2 model, and transmits the trained Client2 model parameters to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module.

[0101] like Figure 4 As shown, step S230 specifically includes:

[0102] S231, after receiving the shallow frozen network of Client1, the line Client2 with the second largest data volume is connected with its own personalized module to form the local model of Client2.

[0103] S232: Build a distillation loss function for the target class (target class knowledge distillation, TCKD) and non-target class (non-target class knowledge distillation, NCKD) between the Client1 model, and perform knowledge distillation based on this to complete the local model training of Client2.

[0104] S233, encrypt the Client2 model parameters and upload them to the central server.

[0105] S234. The central server sends the Client2 model parameters to the remaining low priority lines (Client3, Client4).

[0106] S240, line Clienti uses Clienti local data to train the Clienti model, and transmits the trained Clienti model parameters to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

[0107] Taking four lines as an example, for line Client3, step S240 specifically includes:

[0108] S241, after receiving the shallow frozen network of Client1, the line Client3 with the third largest data volume is connected with its own personalized module to form the local model of Client3.

[0109] S242. Based on the models of Client1 and Client2, Client3 constructs a one-to-one corresponding distillation loss function, and implements model training through a multi-task loss weight adaptive adjustment strategy based on grad-norm to transfer knowledge to the Client3 model.

[0110] S243, encrypt the Client3 model parameters and upload them to the central server.

[0111] S244. The central server sends the model parameters of Client 3 to the remaining low priority lines (Client 4).

[0112] For line Client3, step S240 specifically includes:

[0113] S245: After receiving the shallow frozen network of Client1, the line Client4 with the least amount of data is connected to its own personalized module to form the local model of Client4.

[0114] S246. Based on the models of Client1, Client2 and Client3, Client4 constructs a one-to-one corresponding distillation loss function, implements model training through the grad-norm-based multi-task loss weight adaptive adjustment strategy, transfers knowledge to the Client4 model, and finally realizes the step-by-step federated distillation learning of multi-line failures in the data imbalance scenario.

[0115] In this embodiment, in each round of learning, after the line serving as the teacher completes local training, part of its model will be frozen and shared with other lines. The frozen model does not participate in the local training of other lines, but only serves as a baseline model to provide support for local training, thereby avoiding poor training results caused by training from scratch.

[0116] At the same time, each line will connect its own personalized model after the shared frozen model. According to the characteristics of its own data, based on the designed multi-loss weight calculation method, adaptive training is performed through the distillation strategy, so that the model converges on the local data set to ensure the performance of the overall model. Figure 5 shown.

[0117] In some embodiments, line Client2 uses Client2 local data to train the Client2 model, including:

[0118] The shallow frozen network received from the central server does not participate in the training; a distillation loss function is constructed for the target class and non-target class using the model parameters of Client1, and knowledge distillation is performed based on the distillation loss function to complete the local model training of Client2.

[0119] Each model training is different from the conventional distillation method. The framework distills the knowledge between the two lines each time, and uses two types of losses based on the target class and non-target class to jointly represent the distillation loss.

[0120] Among them, the classification probability p based on the target class t for:

[0121]

[0122] Among them, C is the number of categories, t represents the target category, z j represents the logarithmic value of the jth class, z t represents the number of pairs of target classes.

[0123] Based on the classification probability p of the non-target class \t for:

[0124]

[0125] Among them, z k Represents the logarithmic value of the non-target class.

[0126] After removing the target class, the classification probability of other classes for:

[0127]

[0128] Therefore, when Kullback-Leibler (KL) divergence is used to measure the distillation loss between lines T and S, it is as follows:

[0129]

[0130] Among them, KD represents the distillation loss (Knowledge Distillation), KL represents the KL (Kullback-Leibler) divergence calculation, and p T represents the classification probability distribution of route T, p S represents the classification probability distribution of route S, represents the classification probability of target class t in line T, represents the classification probability of target class t in line S, represents the classification probability of non-target class i in line T, represents the classification probability of non-target class i in route S;

[0131] because The above formula can be rewritten to obtain the distillation loss function as follows:

[0132]

[0133] Among them, b represents the binary classification scenario, TCKD is the target class knowledge distillation, and NCKD is the non-target class knowledge distillation. is the classification probability of target class t in line T, is the sum of the classification probabilities of non-target classes in line T, is the classification probability of the remaining categories i in line T after removing the target class t, is the classification probability of the remaining categories i in line S after removing the target class t, is the sum of the classification probabilities of non-target classes in line T, is the sum of the classification probabilities of non-target classes in line S, is the classification probability distribution of line T excluding target class t, is the classification probability distribution of line S without target class t, b T and b S They represent the binary classification scenarios of route T and route S respectively;

[0134] TCKD can provide key knowledge about the distillation target, while NCKD focuses more on the potential connections between non-target categories. By combining the two, the knowledge transfer between multiple lines can be effectively characterized.

[0135] In particular, in order to better improve the adaptability between the two, an additional weight value is assigned to TCKD as shown below:

[0136] DKD=α·TCKD+β·NCKD

[0137] Among them, DKD represents the coupled distillation loss (Decoupled Knowledge Distillation), α is the adjustment parameter for balancing TCKD and NCKD, and β is used to represent

[0138] In some embodiments, line Clienti uses Clienti local data to train the Clienti model, including: the shallow frozen network received from the central server does not participate in the training, and the line model parameters with a priority higher than the line Clienti are obtained from the central server; a one-to-one corresponding distillation loss function is constructed between the Clienti model and each line model parameter with a priority higher than the line Clienti, and the weight of the distillation loss function is dynamically adjusted to obtain the overall loss function of the Clienti model, and the Clienti model is trained using the overall loss function.

[0139] It should be noted that the method of constructing a one-to-one corresponding distillation loss function between the lienti model and each line model parameter having a priority higher than the line Clienti can adopt two types of losses based on the target class and the non-target class to jointly represent the distillation loss. The specific calculation process has been explained in the previous embodiment and will not be repeated here.

[0140] In the above-designed architecture, multiple distillations are required between lines to achieve model optimization and adjustment, and the loss function obtained by each knowledge distillation will participate in the target line training, which will result in multiple distillation losses. By dynamically adjusting the weights of these loss functions, the effectiveness of distillation between different lines on model training is treated differently, and the overall accuracy of the model is further improved. The overall framework is as follows Figure 6 .

[0141] The overall loss function is designed as follows:

[0142]

[0143] Among them, for the i-th line in the federated task (i=1,2,L,N), represents the local multi-classification cross entropy loss function value, represents the bi-norm value of the DKD loss of the t-th distillation task, represents the average bi-norm value of DKD loss of all tasks, r t Indicates the loss training completion of the t-th distillation task. λ is a hyperparameter that adjusts the local model training and the knowledge transfer of other parties. The item where 1-λ is located can realize the knowledge distillation between multiple lines. μ is a hyperparameter that controls the joint training rate between multiple tasks. The item where it is located can balance the overall loss and play a role in achieving equal training rate.

[0144] In each distillation task, It will be set to a fixed constant and will not participate in the back propagation of gradients during training, preventing the loss function from drifting falsely to zero. At the same time, the training objectives can be aggregated to the local loss function of each party.

[0145] The local cross entropy loss value of the i-th line is:

[0146]

[0147] Among them, C represents the number of categories, y k Indicates the true label corresponding category, Represents the probability predicted by the softmax function;

[0148] The loss training completion of the t-th distillation task r t for:

[0149] r t =L t / L avg

[0150] Among them, L t is the average cross entropy loss between lines, L avg is the average loss of all lines, r t The smaller the value, the higher the training completion.

[0151] The diagnostic accuracy of each line in the experiment is shown in Table 3. It can be concluded that each line can use the knowledge of other lines to further improve the diagnostic performance of the local model and avoid the leakage of local data.

[0152] Table 3. Fault diagnosis accuracy of train bogies on each line

[0153]

[0154] The multi-line bogie fault diagnosis based on federated learning proposed in the embodiment of the present application has a complete diagnostic process. Based on the step-by-step federated distillation architecture, the problem of safe joint modeling of multi-line high-speed train bogies in data imbalance scenarios is solved. The step-by-step federated distillation architecture has been redesigned on the basic process of federated learning, and the step-by-step distillation strategy is used to improve the performance of small sample line models. For the multiple fault diagnosis distillation tasks in the proposed architecture, the loss value is decoupled into target class loss and non-target class loss and further reconstructed. Secondly, for the multiple distillation losses output by multiple tasks, an adaptive adjustment strategy for multiple loss weights is proposed to coordinate multiple distillation tasks to achieve efficient training.

[0155] In order to reduce the training cost of the fault diagnosis model of the proposed framework, each line shares the basic shallow network based on model migration and uses the frozen module to train the local model. Other lines can draw knowledge from the pre-trained shallow model and customize the personalized deep network to improve the adaptability of the fault diagnosis model and conduct personalized exploration of the local line fault data characteristics. This method not only maintains good diagnostic performance, but also significantly reduces the computational cost of model training.

[0156] The proposed framework shows satisfactory fault diagnosis performance on high-speed train bogies. The main reason is that the framework implements multiple step-by-step distillations for each line, which can fully extract the effective feature knowledge of other lines to the greatest extent. The overall diagnosis accuracy of train bogies on each line is higher than 92%, which is more suitable for fault diagnosis tasks in data imbalance scenarios and meets the diagnosis requirements.

[0157] Figure 7 The present invention also provides a multi-line bogie fault diagnosis device based on federated learning. Figure 7 As shown, the multi-line bogie fault diagnosis device based on federated learning includes a central server 701 and multiple lines 702, and the central server 701 and the multiple lines 702 are all signal-connected:

[0158] The central server 701 is used to sort the multiple lines according to the data volume, and obtain a multiple line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, wherein Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, n≥3;

[0159] Line Client1 is used to build a deep neural network locally, and use the local data of Client1 to train the deep neural network, freeze the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmit the shallow frozen network and Client1 model parameters to the central server;

[0160] Line Client2 is used to train the Client2 model using the local data of Client2, and the trained Client2 model parameters are transmitted to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module;

[0161] Line Clienti is used to train the Clienti model using Clienti local data, and the trained Clienti model parameters are transmitted to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

[0162] In some embodiments, multiple lines are sorted according to the amount of data to obtain a priority-sorted multiple line sequence Client=(Client1, Client2...Clienti...Clientn), where Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, where n≥3;

[0163] Line Client1 builds a deep neural network locally, and uses Client1 local data to train the deep neural network. It freezes the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmits the shallow frozen network and Client1 model parameters to the central server.

[0164] Line Client2 trains the Client2 model using the local data of Client2, and transmits the trained Client2 model parameters to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module;

[0165] Line Clienti trains the Clienti model using Clienti local data, and transmits the trained Clienti model parameters to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

[0166] In some embodiments, line Client2 uses Client2 local data to train the Client2 model, including:

[0167] The shallow frozen network received from the central server does not participate in training;

[0168] A distillation loss function is constructed for the target class and non-target class using the model parameters of Client1, and knowledge distillation is performed based on the distillation loss function to complete the local model training of Client2.

[0169] In some embodiments, line Client2 determines the distillation loss function by the following method:

[0170] The classification probability p based on the target class is calculated by the following formula t :

[0171]

[0172] Among them, C is the number of categories, t represents the target category, z j represents the logarithmic value of the jth class, z trepresents the logarithm of the target class;

[0173] The classification probability p based on the non-target class is calculated by the following formula \t :

[0174]

[0175] Among them, z k Represents the logarithmic value of the non-target class;

[0176] After removing the target class, the classification probability of other classes for:

[0177]

[0178] The KL divergence is used to measure the distillation loss between lines T and S:

[0179]

[0180] Among them, KD represents distillation loss, KL represents KL divergence calculation, and p T represents the classification probability distribution of route T, p S represents the classification probability distribution of route S, represents the classification probability of target class t in line T, represents the classification probability of target class t in line S, represents the classification probability of non-target class i in line T, represents the classification probability of non-target class i in route S;

[0181] make The distillation loss function is expressed as:

[0182]

[0183] Among them, b represents the binary classification scenario, TCKD is the target class knowledge distillation, and NCKD is the non-target class knowledge distillation. is the classification probability of target class t in line T, is the sum of the classification probabilities of non-target classes in line T, is the classification probability of the remaining categories i after removing the target class t in line T, is the classification probability of the remaining categories i in line S after removing the target class t, is the sum of the classification probabilities of non-target classes in line T, is the sum of the classification probabilities of non-target classes in line S, is the classification probability distribution of line T excluding target class t, is the classification probability distribution of line S without target class t, b T and b SThey represent the binary classification scenarios of route T and route S respectively;

[0184] An additional weight value is given to TCKD through the following formula:

[0185] DKD=α·TCKD+β·NCKD

[0186] Among them, DKD represents the coupled distillation loss, α is the adjustment parameter for balancing TCKD and NCKD, and β is used to represent

[0187] In some embodiments, line Clienti trains Clienti model using Clienti local data, including:

[0188] The shallow frozen network received from the central server does not participate in training;

[0189] Obtain line model parameters with a higher priority than line Clienti from the central server;

[0190] A one-to-one corresponding distillation loss function is constructed between the Clienti model and the line model parameters with a higher priority than the line Clienti, and the weight of the distillation loss function is dynamically adjusted to obtain the overall loss function of the Clienti model, and the Clienti model is trained using the overall loss function.

[0191] In some embodiments, the overall loss function is expressed as:

[0192]

[0193] Among them, for the i-th line in the federated task (i=1,2,L,N), represents the local multi-classification cross entropy loss function value, represents the bi-norm value of the DKD loss of the t-th distillation task, represents the average bi-norm value of DKD loss for all tasks, r t represents the loss training completion of the t-th distillation task, λ is a hyperparameter that adjusts the local model training and the knowledge transfer from other parties, the item where 1-λ is located is used to realize the knowledge distillation between multiple lines, μ is a hyperparameter that controls the joint training rate between multiple tasks, and the item where it is located is used to balance the overall loss, and M is the total number of distillation tasks;

[0194] In each distillation task, Set to a fixed constant and do not participate in the back propagation of gradients during training;

[0195] The local cross entropy loss value of the i-th line is:

[0196]

[0197] Among them, C represents the number of categories, y k Indicates the true label corresponding category, Represents the probability predicted by the softmax function;

[0198] The loss training completion of the t-th distillation task r t for:

[0199] r t =L t / L avg

[0200] Among them, L t is the average cross entropy loss between lines, L avg is the average loss of all lines, r t The smaller the value, the higher the training completion.

[0201] In some embodiments, the device further comprises a data acquisition module, the data acquisition module being configured to:

[0202] Obtain normal and fault data of high-speed train bogies on different lines; wherein, the normal and fault data of high-speed train bogies include vibration signals under normal working conditions and vibration signals under fault working conditions, and the fault working conditions include AS air leakage, LD oil leakage, AD oil leakage, LD oil leakage + AD oil leakage, AS air leakage + LD oil leakage and AS air leakage + AD oil leakage.

[0203] The multi-line bogie fault diagnosis device based on federated learning provided in the embodiment of the present application can be used to execute the technical solution of the multi-line bogie fault diagnosis method based on federated learning in the above-mentioned embodiment. Its implementation principle and technical effect are similar and will not be repeated here.

[0204] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 8 As shown, the electronic device may include: a processor 81 and a memory 82 , wherein the processor 81 and the memory 82 may communicate with each other; illustratively, the processor 81 and the memory 82 communicate with each other via a communication bus 83 .

[0205] The processor 81 executes the computer execution instructions stored in the memory 82, so that the processor 81 executes the scheme in the above embodiment. The processor 81 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gates or transistor logic devices, and discrete hardware components.

[0206] The communication bus 83 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The transceiver is used to realize the communication between the database access device and other computers (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory.

[0207] The electronic device provided in the embodiment of the present application may be the terminal device of the above embodiment.

[0208] An embodiment of the present application also provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the technical solution of the multi-line bogie fault diagnosis method based on federated learning in the above embodiment.

[0209] An embodiment of the present application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When at least one processor executes the computer program, it can implement the technical solution of the multi-line bogie fault diagnosis method based on federated learning in the above embodiment.

[0210] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0211] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to implement the solution of this embodiment.

[0212] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The above-mentioned module-composed unit can be implemented in the form of hardware or in the form of hardware plus software functional units.

[0213] The above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the methods of various embodiments of the present application.

[0214] It should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented as being executed by a hardware processor, or may be implemented by a combination of hardware and software modules in the processor.

[0215] The memory may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0216] The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0217] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0218] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic control unit or a main control device.

[0219] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0220] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-line bogie fault diagnosis method based on federated learning, characterized in that: include: Sort multiple lines according to the amount of data, and obtain a multi-line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, where Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, n≥3; Line Client1 constructs a deep neural network locally, and trains the deep neural network using Client1 local data, freezes the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmits the shallow frozen network and Client1 model parameters to the central server; Line Client2 trains the Client2 model using the Client2 local data, and transmits the trained Client2 model parameters to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module; Line Clienti trains the Clienti model using Clienti local data, and transmits the trained Clienti model parameters to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

2. The method according to claim 1, characterized in that Line Client2 uses Client2 local data to train the Client2 model, including: The shallow frozen network received from the central server does not participate in training; A distillation loss function is constructed for the target class and the non-target class using the model parameters of Client1, and knowledge distillation is performed based on the distillation loss function to complete the local model training of Client2.

3. The method according to claim 2, characterized in that The distillation loss function is determined by the following method: The classification probability p based on the target class is calculated by the following formula t : Among them, C is the number of categories, t represents the target category, z j represents the logarithmic value of the jth class, z t represents the logarithm of the target class; The classification probability p based on the non-target class is calculated by the following formula \t : Among them, z k Represents the logarithmic value of the non-target class; After removing the target class, the classification probability of other classes for: The KL divergence is used to measure the distillation loss between lines T and S: Among them, KD represents distillation loss, KL represents KL divergence calculation, and p T represents the classification probability distribution of route T, p S represents the classification probability distribution of route S, represents the classification probability of target class t in line T, represents the classification probability of target class t in line S, represents the classification probability of non-target class i in line T, represents the classification probability of non-target class i in route S; make The distillation loss function is expressed as: Among them, b represents the binary classification scenario, TCKD is the target class knowledge distillation, and NCKD is the non-target class knowledge distillation. is the classification probability of target class t in line T, is the sum of the classification probabilities of non-target classes in line T, is the classification probability of the remaining categories i after removing the target class t in line T, is the classification probability of the remaining categories i in line S after removing the target class t, is the sum of the classification probabilities of non-target classes in line T, is the sum of the classification probabilities of non-target classes in line S, is the classification probability distribution of line T excluding target class t, is the classification probability distribution of line S without target class t, b T and b S They represent the binary classification scenarios of route T and route S respectively; An additional weight value is given to TCKD through the following formula: DKD=α·TCKD+β·NCKD Among them, DKD represents the coupled distillation loss, α is the adjustment parameter for balancing TCKD and NCKD, and β is used to represent 4. The method according to claim 1, characterized in that: Line Clienti uses Clienti local data to train the Clienti model, including: The shallow frozen network received from the central server does not participate in training; Obtain line model parameters with a higher priority than line Clienti from the central server; A one-to-one corresponding distillation loss function is constructed between the Clienti model and each line model parameter having a higher priority than the line Clienti, and the weight of the distillation loss function is dynamically adjusted to obtain an overall loss function of the Clienti model, and the Clienti model is trained using the overall loss function.

5. The method according to claim 4, characterized in that The overall loss function is expressed as: Among them, for the i-th line in the federated task (i=1,2,L,N), represents the local multi-classification cross entropy loss function value, represents the bi-norm value of the DKD loss of the t-th distillation task, represents the average bi-norm value of DKD loss for all tasks, r t represents the loss training completion of the t-th distillation task, λ is a hyperparameter that adjusts the local model training and the knowledge transfer from other parties, the item where 1-λ is located is used to realize the knowledge distillation between multiple lines, μ is a hyperparameter that controls the joint training rate between multiple tasks, and the item where it is located is used to balance the overall loss, and M is the total number of distillation tasks; In each distillation task, Set to a fixed constant and do not participate in the back propagation of gradients during training; The local cross entropy loss value of the i-th line is: Among them, C represents the number of categories, y k Indicates the true label corresponding category, Represents the probability predicted by the softmax function; The loss training completion of the t-th distillation task r t for: r t =L t / L avg Among them, L t is the average cross entropy loss between lines, L avg is the average loss of all lines, r t The smaller the value, the higher the training completion.

6. The method according to claim 1, characterized in that Before sorting the multiple lines according to the data volume, the method further includes: Obtain normal and fault data of high-speed train bogies on different lines; wherein, the normal and fault data of high-speed train bogies include vibration signals under normal working conditions and vibration signals under fault conditions, and the fault conditions include AS air leakage, LD oil leakage, AD oil leakage, LD oil leakage + AD oil leakage, AS air leakage + LD oil leakage and AS air leakage + AD oil leakage.

7. A multi-line bogie fault diagnosis device based on federated learning, characterized in that: It includes a central server and multiple lines, and the central server and the multiple lines are all signal connected: The central server is used to sort the multiple lines according to the amount of data, and obtain a multi-line sequence Client=(Client1, Client2...Clienti...Clientn) sorted by priority, wherein Client1 represents the line with the highest priority, Clienti represents the i-th line, Clientn represents the line with the lowest priority, and n represents the total number of multiple lines, n≥3; Line Client1 is used to construct a deep neural network locally, and train the deep neural network using Client1 local data, freeze the shallow network of the trained deep neural network to obtain a shallow frozen network, and transmit the shallow frozen network and Client1 model parameters to the central server; Line Client2 is used to train the Client2 model using the local data of Client2, and the parameters of the Client2 model obtained by training are transmitted to the central server; wherein the Client2 model includes the shallow frozen network received from the central server and the first personalized module; Line Clienti is used to train the Clienti model using Clienti local data, and the trained Clienti model parameters are transmitted to the central server; wherein the Clienti model includes a shallow frozen network received from the central server and an i-1th personalized module, where i is a natural number and 3≤i≤n.

8. An electronic device, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 6 when executed by a processor.

10. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 6 when being executed by a processor.

Images